Wind speed and pressure estimation device and wind speed and pressure estimation method

Abstract

To provide a wind velocity wind pressure estimation device and a wind velocity wind pressure estimation method capable of estimating wind velocity distribution and wind pressure distribution in a short time, simply and in detail.SOLUTION: A wind velocity wind pressure estimation device includes a wind velocity wind pressure estimating unit 21 that: includes a learned model 25 for performing deep learning, as learning data 2, of building information data 3 representing a three-dimensional shape of a building and its surroundings, and wind velocity distribution data 5 and wind pressure distribution data 6 in a three-dimensional space as wind direction data 4 and teacher data corresponding to the building information data 3 and the wind direction data 4; estimates wind velocity distribution and wind pressure distribution by inputting building information data 31 and wind direction data 32 as estimation objects of the wind velocity distribution and the wind pressure distribution into the learned model 25; and outputs an estimation result 40 based on the estimated wind velocity distribution and wind pressure distribution.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a wind speed and pressure estimation device and a wind speed and pressure estimation method. [Background technology]

[0002] Conventionally, wind environments such as wind around high-rise buildings have been evaluated. The wind environment is sometimes evaluated as wind speed distribution and wind pressure distribution by, for example, wind tunnel experiments or computational fluid dynamics. In wind tunnel experiments, wind is blown onto a model that reproduces a building to be evaluated, and the wind speed around the model is measured with sensors attached to the model, thereby evaluating the wind speed around the building. Patent Document 1 discloses a wind speed measurement device used in wind tunnel experiments. In the wind tunnel experiment described above, wind speed and wind pressure are measured physically by actually blowing wind onto a model, so highly reliable evaluation results can be obtained. However, wind tunnel experiments require large wind tunnel equipment, and building models takes a lot of time and money. Furthermore, since sensors have a certain size, the number that can be installed on the model is limited, and therefore wind speeds cannot be obtained at all points on or around the model, and the density of points at which evaluation results can be obtained tends to be sparse. Furthermore, in wind tunnel experiments, one method for measuring wind speed is to use PIV (Particle Image Velocimetry). Another method for measuring wind pressure is to apply pressure-sensitive paint to the surface of a model and observe changes in its luminescence intensity. However, these methods are limited to obtaining measured and observed values ​​within a two-dimensional plane, making it difficult to obtain measured and observed values ​​in three dimensions for each position in three-dimensional space.

[0003] In computational fluid analysis, the building to be evaluated is modeled on a computer, and an approximate solution for wind speed is calculated by applying equations related to discretized and modeled fluid motion to the model. Patent Document 2 discloses a wind environment prediction method based on the results of fluid analysis. Because computational fluid analysis can be performed using a computer, it does not require the large-scale facilities required for wind tunnel testing, and building modeling can be performed in a shorter time and at lower cost than the production of models required for wind tunnel testing.In addition, the density of points at which evaluation results can be obtained can be higher than with wind tunnel testing. However, advanced knowledge and know-how are required for analysis preparation work such as building modeling and mesh generation, and analyses of urban areas generally require large-scale 3D calculations, which take a considerable amount of time. A typical calculation method for this type is LES (Large Eddy Simulation), which can calculate fluctuations in wind speed and wind pressure as well as maximum and minimum values ​​over time, but LES in particular requires dividing the space into extremely fine grids, which places a heavy burden on the calculations and makes it unrealistic.

[0004] In addition to the wind tunnel experiments and computational fluid analysis mentioned above, attempts have been made in recent years to evaluate wind environments using neural networks. When using a neural network, an evaluation device can be realized using a general-purpose computer, eliminating the need for large-scale equipment such as wind tunnel experiments. Furthermore, once the neural network has completed its learning, it is possible to output the wind environment simply by inputting simple data, so advanced knowledge and know-how, such as with computational fluid analysis, are not generally required. Furthermore, calculation times are generally faster than with computational fluid analysis. For example, Patent Document 3 discloses a wind speed distribution estimation device around a building, which is related to the evaluation of wind speed distribution using a neural network as described above. The wind speed distribution estimation device of Patent Document 3 includes an input means for inputting a height image representing height information to be estimated, a shape image representing cross-sectional shape information of the building, and a wind direction image representing wind direction information, and a wind speed distribution image creation means for creating a wind speed distribution image that is an image representing the wind speed distribution around the building. The wind speed distribution image creation means is composed of a convolutional neural network that uses the height image, shape image, and wind direction image as input images and the wind speed distribution image as an output image. In Patent Document 3, the wind speed distribution image is used as an input for creating a wind speed contour map. In other words, in the wind speed distribution image in Patent Document 3, the pixel value of each pixel represents the wind speed at the point corresponding to that pixel.

[0005] In the wind speed distribution estimation device of Patent Document 3, as described above, each pixel value represents the wind speed at the corresponding point, and therefore information about the wind pressure at each point cannot be obtained. Furthermore, the output of the wind speed distribution estimation device is an image that can only contain two-dimensional information, so as with wind tunnel experiments, it is difficult to estimate the wind speed in detail for each position in three-dimensional space. It is desirable to be able to evaluate wind speed and pressure simply, quickly, and in more detail. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] Japanese Patent Application Laid-Open No. 2014-48120 [Patent Document 2] Japanese Patent Application Publication No. 2018-165884 [Patent Document 3] Japanese Patent Application Laid-Open No. 2018-4568 Summary of the Invention [Problem to be solved by the invention]

[0007] The problem to be solved by the present invention is to provide a wind speed and pressure estimation device and a wind speed and pressure estimation method that can estimate wind speed distribution and wind pressure distribution easily and in detail in a short time. [Means for solving the problem]

[0008] In order to solve the above-mentioned problems, the present invention employs the following means: Specifically, the present invention provides a wind speed and pressure estimation device that estimates wind speed and wind pressure distributions around a building, the device comprising a trained model that has been deep-learned using, as training data, building information data representing the three-dimensional shape of a building and its surroundings, wind direction data, and wind speed distribution data and wind pressure distribution data in three-dimensional space as teacher data corresponding to the building information data and the wind direction data, and a wind speed and wind pressure estimation unit that inputs the building information data and wind direction data from which the wind speed distribution and wind pressure distribution are to be estimated into the trained model to estimate the wind speed distribution and wind pressure distribution and outputs an estimation result based on this, wherein the wind speed distribution data and the wind speed distribution estimated by the trained model each store wind speed information corresponding to each of a plurality of mutually perpendicular axial directions at each position in the three-dimensional space, and the wind pressure distribution data and the wind pressure distribution estimated by the trained model each store wind pressure information at each position in the three-dimensional space. According to the above configuration, the trained model is deep-trained using building information data, wind direction data, and wind speed distribution data and wind pressure distribution data in three-dimensional space as training data corresponding to the building information data and wind direction data. When the building information data and wind direction data for which wind speed distribution and wind pressure distribution are to be estimated are input to this trained model, the trained model estimates the wind speed distribution and the wind pressure distribution. Here, the wind speed distribution data and the wind speed distribution estimated by a trained model that has undergone deep learning to output values ​​close to this wind speed distribution data as training data each store wind speed information corresponding to each of multiple mutually orthogonal axial directions at each position in three-dimensional space. In other words, the trained model can estimate wind speeds corresponding to multiple axial directions at each position in three-dimensional space. Furthermore, by combining the wind speeds in each of the multiple axial directions at each position, it is possible to calculate not only the wind speed but also the wind direction. Furthermore, the wind pressure distribution data and the wind pressure distribution estimated by the trained model, which has undergone deep learning to output values ​​close to this wind pressure distribution data as training data, each store wind pressure information at each position in three-dimensional space. In other words, the trained model can estimate wind pressure at each position in three-dimensional space. In particular, the trained model configured as described above estimates wind speed distribution and wind pressure distribution at each position in three-dimensional space together. In other words, the trained model can estimate results with continuity in each axial direction in three-dimensional space. Therefore, the estimation results are more accurate and detailed than methods that estimate two-dimensional information. Furthermore, in the above configuration, wind speed distribution and wind pressure distribution are estimated using a deep learning trained model, which can be achieved in a short time and with simpler processing than when using wind tunnel experiments or computational fluid analysis. Therefore, it is possible to realize a wind speed and wind pressure estimation device that can estimate wind speed distribution and wind pressure distribution easily and in detail in a short time.

[0009] In one aspect of the present invention, the building information data is modeled as three-dimensional data consisting of two values, with building parts being 0 and non-building parts being 1, and the parts of the wind speed distribution data and the wind pressure distribution data that correspond to the building parts are set to a value of 0, and during learning, the trained model is deep-trained by comparing the results of multiplying the values ​​of the estimated wind speed distribution and wind pressure distribution at each position in the three-dimensional space by the value of the building information data corresponding to that position with the wind speed distribution data and the wind pressure distribution data. During learning, the trained model is basically deep-learned by comparing each of the estimated wind speed distribution and wind pressure distribution with wind speed distribution data and wind pressure distribution data as training data, and reflecting the differences. In the above configuration, the values ​​of the wind speed distribution data and wind pressure distribution data corresponding to building parts are set to 0. The building information data is modeled as binary three-dimensional data, with building parts set to 0 and non-building parts set to 1. During training, the trained model performs deep learning by comparing the values ​​of the estimated wind speed distribution and wind pressure distribution at each position in three-dimensional space, multiplied by the building information data value corresponding to that position, with the wind speed distribution data and wind pressure distribution data. That is, in the results of multiplying the estimated wind speed distribution and wind pressure distribution at training by the building information data value, which are the basis for deep learning and the objects for which differences are sought through comparison, and in the wind speed distribution data and wind pressure distribution data, both the building parts are set to 0. Therefore, even if the trained model estimates a value other than 0 for the building parts as wind speed or wind pressure during training, this value is multiplied by 0 when compared with the teacher data, i.e., the wind speed distribution data and wind pressure distribution data, resulting in a value of 0. Therefore, no difference occurs between the teacher data and the trained model, and as a result, it is not reflected in the training of the trained model. This configuration prevents the trained model from excessively learning features related to the shape of buildings, etc., and as a result, it can focus on learning features related to wind speed and wind pressure in non-building areas. This improves the learning efficiency of the trained model and further increases the estimation accuracy of wind speed distribution and wind pressure distribution.

[0010] The present invention also provides a wind speed and pressure estimation method for estimating wind speed distribution and wind pressure distribution around a building, comprising a step of inputting building information data and wind direction data representing the three-dimensional shape of a building and its surroundings, and wind speed distribution data and wind pressure distribution data in three-dimensional space as teacher data corresponding to the building information data and the wind direction data into a trained model that has been deep-learned as learning data, to estimate the wind speed distribution and wind pressure distribution, and outputting an estimation result based on this, wherein the wind speed distribution data and the wind speed distribution estimated by the trained model each store wind speed information corresponding to each of a plurality of mutually perpendicular axial directions at each position in the three-dimensional space, and the wind pressure distribution data and the wind pressure distribution estimated by the trained model each store wind pressure information at each position in the three-dimensional space. According to the above configuration, it is possible to realize a wind speed and wind pressure estimation method that can estimate wind speed distribution and wind pressure distribution easily and in detail in a short time. [Effects of the Invention]

[0011] According to the present invention, it is possible to provide a wind speed and pressure estimation device and a wind speed and pressure estimation method that can estimate wind speed distribution and wind pressure distribution easily and in detail in a short time. [Brief explanation of the drawings]

[0012] [Figure 1] 1 is a block diagram of a wind speed and wind pressure estimation device according to an embodiment of the present invention. [Figure 2] FIG. 2 is a schematic explanatory diagram of a machine learning machine in the wind speed and pressure estimation device. [Figure 3]FIG. 2 is an explanatory diagram of building information data that is input to the machine learning machine. [Figure 4] FIG. 2 is an explanatory diagram of wind direction data that is input to the machine learning machine. [Figure 5] FIG. 10 is an explanatory diagram of wind speed distribution X-component data that is input to the machine learning machine. [Figure 6] FIG. 10 is an explanatory diagram of wind speed distribution Y component data that is input to the machine learning machine. [Figure 7] FIG. 10 is an explanatory diagram of wind speed distribution Z component data that is input to the machine learning machine. [Figure 8] FIG. 2 is an explanatory diagram of wind pressure distribution data that is input to the machine learning machine. [Figure 9] FIG. 1 is a schematic explanatory diagram of a trained model after the machine learning machine has completed training. [Figure 10] FIG. 10 is an explanatory diagram of a result of wind speed distribution synthesis. [Figure 11] FIG. 10 is an explanatory diagram of the wind pressure distribution estimation result. [Figure 12] 10 is a flowchart of a deep learning process for a wind speed and pressure estimation method using the wind speed and pressure estimation device. [Figure 13] 10 is a flowchart showing a wind speed and wind pressure estimation method using the wind speed and wind pressure estimation device. DETAILED DESCRIPTION OF THE INVENTION

[0013] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. FIG. 1 is a block diagram of a wind speed distribution estimation device according to this embodiment. As described above, the wind speed and pressure estimation device 1 in this embodiment is a device that estimates wind speed distribution and wind pressure distribution around an arbitrary building. The wind speed and pressure estimation device 1 is an information processing device such as a personal computer. The wind speed and pressure estimation device 1 includes a learning unit 20, a wind speed and pressure estimation unit 21, a learned model parameter storage unit 22, and a result synthesis unit 23. Of these components of the wind speed and pressure estimation device 1, the learning unit 20, wind speed and pressure estimation unit 21, and result synthesis unit 23 may be software or programs executed by, for example, a CPU in the information processing device. The trained model parameter storage unit 22 may be realized by a storage device such as a semiconductor memory or a magnetic disk provided inside or outside the information processing device. The learning unit 20 may be processed by, for example, a GPU (Graphics Processing Unit).

[0014] As will be explained later, when building information data 31 representing the three-dimensional shape of a building and its surroundings and wind direction data 32 are input, wind speed and pressure estimation unit 21 estimates the corresponding wind speed distribution and wind pressure distribution and outputs estimation results 40 based on the data. To perform this estimation effectively, wind speed and pressure estimation unit 21 is equipped with a trained model 25 generated by machine learning using a machine learning device 24 provided in learning unit 20. More specifically, learning unit 20 inputs training data 2 to machine learning device 24 and performs machine learning to generate trained model parameters related to the estimation of wind speed distribution and wind pressure distribution. In other words, machine learning device 24 is a trained model 25 during learning, i.e., in the middle of learning. That is, the wind speed and pressure estimation device 1 performs two main operations: learning wind speed and pressure distribution, and estimating wind speed and pressure distribution. To simplify the explanation, the following will first explain each component of the wind speed and pressure estimation device 1 when learning wind speed and pressure distribution, and then explain the behavior of each component when estimating wind speed and pressure distribution.

[0015] When learning wind speed distribution, the learning unit 20 causes the machine learning device 24 to perform machine learning based on the learning data 2. This machine learning device 24 performs deep learning, thereby generating a trained model 25. Figure 2 is a schematic explanatory diagram of the machine learning device 24. The learning data 2 includes building information data 3, wind direction data 4, and wind speed distribution data 5 and wind pressure distribution data 6 as training data. The building information data 3 represents the locations of buildings and other structures that obstruct the passage of wind within a three-dimensional target region for which wind speed and wind pressure distributions are estimated by the wind speed and wind pressure estimation device 1. The building information data 3 subdivides the target region into multiple mutually orthogonal axial directions X, Y, and Z: an X direction, which is a first axial direction extending within a horizontal plane; a Y direction, which is a second axial direction extending within the horizontal plane and perpendicular to the X direction; and a Z direction, which is a third axial direction perpendicular to the horizontal plane, i.e., perpendicular to both the X and Y directions. In other words, the building information data 3 is voxel data represented as a set of voxels formed by connecting voxels, which are the smallest units constituting the building information data 3, in each of the X, Y, and Z directions. In this embodiment, a value of 0 is set for building portions 3a of the building information data 3, which correspond to structures such as buildings that obstruct wind, and a value of 1 is set for non-building portions 3b, which correspond to other portions through which wind can pass. In this way, the building information data 3 is modeled as three-dimensional data consisting of two values, 0 and 1, to represent the three-dimensional shapes of buildings and their surroundings within the target area. In FIG. 2, only voxels corresponding to building portions 3a and set to a value of 0 are shown, but in reality, the entire target area, including non-building portions 3b, is made up of voxels.

[0016] The wind direction data 4 is data that contains information about wind direction. More specifically, the wind direction data 4 is data that expresses the direction of wind blowing from outside the region toward a building or group of buildings within the region expressed in the building information data 3, as already explained. The wind direction is uniquely determined for all points within the region, regardless of the point within the region in the building information data 3. The wind direction data 4 represents the wind direction as described above, and because the value is unique at every point, it can be regarded as a single vector. When this vector is decomposed into components in the X and Y directions in the building information data 3, the wind direction data 4 includes wind direction X component data 8 that represents the component value in the X direction, and wind direction Y component data 9 that represents the component value in the Y direction. The wind direction X-component data 8 and the wind direction Y-component data 9 are voxel data configured to have the same size as the building information data 3 in each of the X, Y, and Z directions. For example, the wind direction X-component data 8 can be voxel data in which all voxels are uniformly set to values ​​corresponding to the component values ​​in the X direction into which a vector is decomposed. Also, the wind direction Y-component data 9 can be voxel data in which all voxels are uniformly set to values ​​corresponding to the component values ​​in the Y direction into which a vector is decomposed. Note that while the wind direction X-component data 8 and the wind direction Y-component data 9 are each shown two-dimensionally in FIG. 2, in reality, they are voxel data having a three-dimensional structure, as explained above.

[0017] FIG. 4 is an explanatory diagram of wind direction data. In this embodiment, as shown in FIG. 4, in a two-dimensional coordinate system formed by the X and Y directions in the building information data 3, where X and Y each can have a value between 0 and 1, the value of the wind direction data 4 is set as the downstream coordinate when the coordinate (0.5, 0.5) is set as the upstream wind direction. For example, in the case of a westerly wind, if the coordinate (0.5, 0.5) is taken as the upstream in a coordinate system like that shown in Figure 4, the wind direction will be to the right on the page, indicated as R1, so the downstream coordinate will be (1, 0.5), and therefore the component value in the X direction will be 1 and the component value in the Y direction will be 0.5. In the case of a southerly wind, if the coordinate (0.5, 0.5) is taken as the upstream, the wind direction will be to the top of the page, indicated as R2, so the downstream coordinate will be (0.5, 1), and therefore the component value in the X direction will be 0.5 and the component value in the Y direction will be 1. Similarly, in the case of a southwesterly wind, the wind direction will be in the direction indicated as R3, so the downstream coordinate will be (1, 1), and therefore the component value in the X direction will be 1 and the component value in the Y direction will be 1. In the case of a wind from the west-southwest, the wind direction is shown as R4, so the downstream coordinate is (1, 0.75), and therefore the component value in the X direction is 1 and the component value in the Y direction is 0. Furthermore, in the case of a wind from the northeast, the wind direction is shown as R5, so the downstream coordinate is (0, 0), and therefore the component value in the X direction is 0 and the component value in the Y direction is 0. The wind direction X-component data 8 contains the component value in the X direction calculated as described above. The wind direction Y-component data 9 contains the component value in the Y direction calculated as described above.

[0018] The wind speed distribution data 5 is information about the distribution of wind speeds in the area represented in the building information data 3, i.e., data about the wind speeds at various points around the building. The wind speed distribution data 5 is data obtained in response to the building information data 3 and the wind direction data 4, and is uniquely determined based on the building information data 3, which is information about the shape and height of the building, and the wind direction data 4, which is information about the wind blowing into the target area represented in the building information data 3. In particular, in this embodiment, the wind speed distribution data 5 also includes information about the wind direction at each point. To this end, the wind speed distribution data 5 includes wind speed distribution X-component data 10, wind speed distribution Y-component data 11, and wind speed distribution Z-component data 12, which are voxel data in three-dimensional space and each configured to have the same size as the building information data 3, so that the wind speed distribution data 5 can hold information about the wind direction.

[0019] 5, 6, and 7 are explanatory diagrams of wind speed distribution X-component data, wind speed distribution Y-component data, and wind speed distribution Z-component data, respectively. The wind speed distribution X-component data 10, wind speed distribution Y-component data 11, and wind speed distribution Z-component data 12 are voxel data configured to have the same size as the building information data 3 in each of the X, Y, and Z directions. In the wind speed distribution X-component data 10, the value of each voxel is set to a value equivalent to the component value in the X direction when a vector representing the wind direction and wind speed at a point corresponding to the voxel in the target area represented in the building information data 3 is decomposed into components in each of the X, Y, and Z directions in the building information data 3, for given building information data 3 and wind direction data 4. Similarly, the wind speed distribution Y-component data 11 and wind speed distribution Z-component data 12 have the value of each voxel set to a value equivalent to the component value in each of the Y and Z directions. As mentioned above, each of the wind speed distribution X-component data, wind speed distribution Y-component data, and wind speed distribution Z-component data is actually voxel data and has a three-dimensional structure, but in Figures 5, 6, and 7 they are shown as two planes: a horizontal plane corresponding to the ground surface and a plane perpendicular to it. In each of the wind speed distribution X-component data 10, wind speed distribution Y-component data 11, and wind speed distribution Z-component data 12, if the point corresponding to a voxel is a building, i.e., if it corresponds to the building part 3a, the value of the voxel is set to 0.

[0020] As will be described later, the wind speed distribution X-component data 10, wind speed distribution Y-component data 11, and wind speed distribution Z-component data 12 are used as training data when deep learning is performed on the machine learning device 24. The wind speed distribution X-component data 10, wind speed distribution Y-component data 11, and wind speed distribution Z-component data 12 can be obtained by wind tunnel experiments, computational fluid analysis, actual measurements, or other means. In particular, in this embodiment, the results of computational fluid analysis are used because three-dimensional wind speed and wind pressure distributions can be calculated relatively easily. In this way, when obtaining the actual values ​​of the wind speed distribution X-component data 10, wind speed distribution Y-component data 11, and wind speed distribution Z-component data 12, the surrounding environment is stored as building information data 3, and the wind direction at that time is stored as wind direction data 4, corresponding to the wind speed distribution X-component data 10, wind speed distribution Y-component data 11, and wind speed distribution Z-component data 12, respectively. In practice, a large amount of training data is required for the machine learning device 24 to learn. Therefore, for example, the wind speed distribution X-component data 10, wind speed distribution Y-component data 11, and wind speed distribution Z-component data 12 are acquired using not only computational fluid dynamics analysis but also other techniques such as wind tunnel experiments and actual measurements. However, these distribution data are not necessarily obtained under the same wind speed conditions. For example, in the case of wind tunnel experiments, the wind speed of the wind tunnel airflow may vary from case to case, and in the case of computational fluid dynamics analysis, the wind speed applied to the inflow boundary may vary from case to case. In the case of actual measurements, it is easily assumed that the on-site wind speed varies from case to case. For this reason, in this embodiment, the values ​​set for voxels in the voxel data of the wind speed distribution X-component data 10, wind speed distribution Y-component data 11, and wind speed distribution Z-component data 12 are normalized by the wind speed at a predetermined reference height, which is the wind speed reference height.

[0021] For example, a height position higher than the upper limit height in the Z direction of the target area represented in the building information data 3 can be adopted as this reference height. For example, the Building Standards Act and the Building Load Guidelines define ground surface roughness classifications based on the size and density of buildings surrounding a construction site, and define boundary layer heights and power-law exponents according to these classifications. The boundary layer height is the height at which wind speed attenuates due to the influence of surrounding buildings, and wind speeds above the boundary layer height are conveniently considered constant. For example, the Building Load Guidelines stipulate that the boundary layer height is 450 m in areas with scattered mid-rise buildings, 550 m in urban areas dominated by mid-rise buildings, and 650 m in urban areas densely populated by high-rise buildings. Furthermore, the degree of wind speed attenuation is approximated by an exponential function, and this attenuation degree is defined by the power-law exponent. Standardizing wind speeds at boundary layer heights calculated using the power-law exponent corresponding to the ground surface roughness classification makes it possible to match wind speed conditions even when wind speeds differ between the data obtained. More conceptually, the boundary layer height can be considered to be the height at which the wind speed is lower at positions lower than the boundary layer height due to the influence of buildings than at positions higher than the boundary layer height. Therefore, in this embodiment, it is considered that the wind speed at heights lower than the boundary layer height will not generally be higher than at the boundary layer height, and the wind speed is normalized so that the wind speed at the boundary layer height is 1 and the value of each voxel located within the target area represented in the building information data 3 is a value smaller than 1. In this embodiment, the reference height is set to 450 m. However, according to the above-mentioned building load guidelines, the reference height may be set to 550 m in urban areas where mid-rise buildings are the norm, or 650 m in urban areas where high-rise buildings are densely located.

[0022] More specifically, the value of each voxel of the wind speed distribution X-component data 10, the wind speed distribution Y-component data 11, and the wind speed distribution Z-component data 12 is set as follows. First, the set wind speed in the wind tunnel experiment or numerical fluid analysis at the reference height set as above, or the actual measured value in the case of actual measurement, is set as the reference wind speed. Next, in each of the wind speed distribution X-component data 10, wind speed distribution Y-component data 11, and wind speed distribution Z-component data 12, the value of each voxel is set to a value obtained by normalizing the wind speed corresponding to the voxel to a value between 0 and 1. During this normalization, when there is no wind, i.e., when the wind speed is 0, the value is set to 0.5, which is the intermediate value between 0 and 1. Then, when the wind speed is a positive value, i.e., when the wind is blowing in the positive X, Y, or Z direction in each of the wind speed distribution X-component data 10, wind speed distribution Y-component data 11, and wind speed distribution Z-component data 12, the value is normalized to a value greater than 0.5 and less than 1. When the wind speed is a negative value, i.e., when the wind is blowing in the direction opposite to the positive X, Y, or Z direction in each of the wind speed distribution X-component data 10, wind speed distribution Y-component data 11, and wind speed distribution Z-component data 12, the value is normalized to a value less than 0.5 and greater than 0. Such normalization can be achieved by calculating a wind speed ratio by dividing the wind speed at each location by a reference wind speed, then adding 1 to the wind speed ratio and dividing it by 2. For example, if the wind speed in the X direction at a certain point within the target area represented in the building information data 3 is 10 m / s and the reference wind speed is 13 m / s, the wind speed ratio is 10 / 13, so 0.885, which is obtained by adding 1 to this ratio and dividing it by 2, is set as the value of the voxel corresponding to that point in the wind speed distribution X-component data 10. Furthermore, if the wind speed in the X direction at a certain point within the target area represented in the building information data 3 is −5 m / s, i.e., the wind speed is in the opposite direction to the X direction, the wind speed ratio is −5 / 13, and the value obtained by adding 1 to this and dividing by 2, or 0.308, is set as the value of the voxel corresponding to that point in the wind speed distribution X-component data 10. In this way, the wind speed distribution data 5 stores wind speed information corresponding to each of a plurality of axial directions X, Y, and Z that are orthogonal to one another at each position in three-dimensional space.

[0023] The wind pressure distribution data 6 is information about the distribution of wind pressure in the area represented in the building information data 3, i.e., data about wind pressure at each point around the building. The wind pressure distribution data 6 is data obtained in response to the building information data 3 and wind direction data 4, and like the wind speed distribution data 5, when the building information data 3 and wind direction data 4 are given, the wind pressure distribution data 6 is uniquely determined based on these. FIG. 8 is an explanatory diagram of wind pressure distribution data. The wind pressure distribution data 6 is voxel data in three-dimensional space, in which the value of each voxel is set to a value corresponding to the wind pressure at a point corresponding to the voxel in the target area expressed in the building information data 3, for given building information data 3 and wind direction data 4. As described above, the wind pressure distribution data 6 is actually voxel data and has a three-dimensional structure, but in Fig. 8 it is shown as two planes: a horizontal plane corresponding to the ground surface and a plane perpendicular to it. In the wind pressure distribution data 6, if the point corresponding to a voxel is a building, that is, if it corresponds to the building part 3a, the value of the voxel is set to 0.

[0024] As will be explained later, the wind pressure distribution data 6 is used as training data when deep learning is performed on the machine learning machine 24. Like the wind speed distribution data 5, each piece of wind pressure distribution data 6 can be obtained by means of wind tunnel experiments, numerical fluid analysis, actual measurements, etc. When obtaining the actual value of the wind pressure distribution data 6, the surrounding environment is stored as building information data 3, and the direction of the wind at that time is stored as wind direction data 4, each corresponding to the wind pressure distribution data 6. In this embodiment, the value of each voxel of the wind pressure distribution data 6 is standardized using a value of 450 m as the reference wind pressure, similar to the wind speed distribution data 5. If the reference wind pressure cannot be obtained, the velocity pressure, which is a value obtained by multiplying the square of the reference wind speed by 0.61, may be used as the reference wind pressure. In this embodiment, the values ​​of the wind pressure distribution data 6 are also normalized in the same manner as the wind speed distribution data 5. In this way, the wind pressure distribution data 6 stores wind pressure information at each position in three-dimensional space.

[0025] When machine learning is performed on the machine learning device 24, the building information data 3 and the wind direction data 4 are used as inputs to the machine learning device 24. The learning output data 15 output by the machine learning device 24 at this time is compared with the wind speed distribution data 5 and wind pressure distribution data 6 corresponding to the building information data 3 and wind direction data 4, and the machine learning device 24 is trained based on the results of this comparison. In this way, the wind speed distribution data 5 and the wind pressure distribution data 6 are used as correct values ​​in the machine learning, i.e., training data, and the machine learning device 24 is trained to output learning output data 15 that is close to the wind speed distribution data 5 and wind pressure distribution data 6. As already explained, the building information data 3, the wind direction data 4, and the wind speed distribution data 5 and wind pressure distribution data 6 serving as training data are each voxel data. Therefore, the learning output data 15 compared with the wind speed distribution data 5 and wind pressure distribution data 6 is also voxel data. For this reason, in this embodiment, the machine learning device 24 is implemented by a three-dimensional full-layer convolutional network, i.e., a 3D-FCN (Fully Convolutional Network), which is compatible with processing when voxel data is used as input and output. As will be explained below, an FCN does not include a fully connected layer, and a feature map processed and generated in a convolutional layer is directly input to a transposed convolutional layer.

[0026] As shown in Fig. 2, the machine learning machine 24 includes a convolution processing unit 27 and a transposed convolution processing unit 28. The convolution processing unit 27 includes a plurality of convolution layers 27a, 27b, and 27c connected in series. The transposed convolution processing unit 28 includes a plurality of transposed convolution layers 28c, 28b, and 28a connected in series as well. In the schematic diagram of Fig. 2, the number of convolution layers and the number of transposed convolution layers are each shown as three, but the number of these convolution layers and transposed convolution layers is not limited to three. The learning unit 20 inputs the building information data 3 and the wind direction data 4 to the first convolutional layer 27a. As already explained, the building information data 3 and the wind direction X-component data 8 and wind direction Y-component data 9 that make up the wind direction data 4 are each voxel data configured to have the same size in each of the X, Y, and Z directions. In practice, the building information data 3, wind direction X-component data 8, and wind direction Y-component data 9 are integrated as training input data 14, which is a single voxel data, and this training input data 14 is input to the machine learning device 24. The training input data 14 can be integrated, for example, by setting the values ​​of voxels corresponding to the same location in the building information data 3, wind direction X-component data 8, and wind direction Y-component data 9 to be values ​​of different channels within the voxels at the corresponding positions in the training input data 14.

[0027] The convolutional layer 27a includes a predetermined number of filters. The machine learning machine 24 performs convolutional filtering by positioning each filter on the training input data 14 and calculating the sum of the values ​​of each voxel in the training input data 14 within the filter, with weights set in the filter corresponding to the position of the voxel. This calculates the value of one voxel in the convolutional layer 27a. The machine learning machine 24 performs this convolutional filtering while moving the filter on the training input data 14 in predetermined resolution increments, thereby calculating the values ​​of multiple voxels, and arranging these values ​​to generate one piece of voxel data corresponding to the filter. The machine learning machine 24 executes this process for all filters and generates feature maps according to the number of filters. Optionally, batch normalization, pooling, and activation functions are applied to the feature maps. The feature map generated in the convolutional layer 27a becomes voxel data to be input to the next convolutional layer 27b.

[0028] In the convolutional layer 27b, convolutional filtering is performed on the feature maps generated in the convolutional layer 27a, in the same manner as in the convolutional layer 27a. The convolutional layer 27b is provided with a predetermined number of filters, and performs convolutional filtering using these filters, and further performs batch normalization and pooling as necessary, thereby generating a predetermined number of feature maps corresponding to the number of filters. In the convolutional layer 27c, convolutional filtering is performed on the feature maps generated in the convolutional layer 27b. The convolutional layer 27c is provided with a predetermined number of filters, and performs convolutional filtering using these filters, and further performs batch normalization and pooling as necessary, thereby generating a predetermined number of feature maps corresponding to the number of filters. The filter weights in each of the convolution layers 27a, 27b, and 27c are adjusted by machine learning. The feature map generated in the convolution layer 27c is input to the transposed convolution layer 28c of the transposed convolution processing unit 28.

[0029] The transposed convolution processing unit 28 has a structure symmetrical to that of the convolution processing unit 27. That is, the training input data 14 is compressed to a low dimension by the convolution processing unit 27, but the transposed convolution processing unit 28 operates to expand and restore the data from the low-dimensional compressed state. More specifically, output data is generated by passing through a transposed convolution layer 28c that performs a transposed convolution process corresponding to the convolution layer 27c, a transposed convolution layer 28b that performs a transposed convolution process corresponding to the convolution layer 27b, and a transposed convolution layer 28a that performs a transposed convolution process corresponding to the convolution layer 27a, in that order. In this embodiment, the output data of the transposed convolution processing unit 28, i.e., the machine learning machine 24, is training output data 15, which is the current learning stage estimation result of the wind speed distribution and wind pressure distribution in the target area represented in the building information data 3, corresponding to the building information data 3 and wind direction data 4 in the input training input data 14. The training output data 15 is voxel data configured to have the same size as the training input data 14 in each of the X, Y, and Z directions. The training output data 15 is configured so that each voxel has four channels, first to fourth. The first channel of each voxel in the training output data 15 stores the estimation result of the wind speed in the X direction at the position corresponding to the voxel. The second channel of each voxel in the training output data 15 stores the estimation result of the wind speed in the Y direction at the position corresponding to the voxel. The third channel of each voxel in the training output data 15 stores an estimation result of the wind speed in the Z direction at the position corresponding to that voxel. The fourth channel of each voxel in the training output data 15 stores an estimation result of the wind pressure at the position corresponding to that voxel. By individually separating the values ​​of the first, second, third, and fourth channels for each voxel in the training output data 15, it is possible to obtain the wind speed distribution of the X component, the wind speed distribution in the Y direction, the wind speed distribution in the Z direction, and the wind pressure distribution. In this way, the wind speed distribution estimated by the machine learning device 24, which is the trained model 25 during training (i.e., training output data 15), stores wind speed information corresponding to each of multiple mutually perpendicular axial directions X, Y, and Z at each position in three-dimensional space, and the wind pressure distribution estimated by the machine learning device 24 (i.e., training output data 15) stores wind pressure information at each position in three-dimensional space.

[0030] Here, in order to perform machine learning efficiently, in this embodiment, the learning unit 20 multiplies the learning output data 15 estimated by the machine learning device 24 by the building information data 3 to generate the learning estimation result 16. More precisely, the learning unit 20 multiplies each voxel in the learning output data 15 by the value of the voxel at the corresponding position in the building information data 3 to generate the learning estimation result 16. As already explained, in the building information data 3, the value of the voxel corresponding to the building portion 3a is set to 0, and the value of the voxel corresponding to the non-building portion 3b is set to 1. Therefore, by performing this processing, the value of the voxel corresponding to the building portion 3a in the learning estimation result 16 is set to 0, and the value of the voxel corresponding to the non-building portion 3b is set so that the value of the corresponding voxel in the learning output data 15 remains unchanged. As will be explained below, the machine learning machine 24 basically performs deep learning by comparing the learning estimation result 16 with the wind speed distribution data 5 and wind pressure distribution data 6 as teacher data and reflecting the differences. Here, the building portion 3a is set to 0 in each of the learning estimation result 16, the wind speed distribution data 5, and the wind pressure distribution data 6, which are the basis for deep learning and the targets for which differences are found by comparison. For this reason, even if the machine learning machine 24 estimates a value other than 0 for the building portion 3a as the wind speed or wind pressure during learning, when this is compared with the teacher data, i.e., the wind speed distribution data 5 and the wind pressure distribution data 6, the value is multiplied by 0 to make it 0, so no difference occurs with the teacher data, and as a result, it is not reflected in the learning of the machine learning machine 24. In other words, by using the learning estimation results 16 that have undergone the above-described processing for deep learning, the machine learning device 24 does not need to overlearn features related to the shape of the building, etc., and as a result, it can concentrate on learning only the features related to the wind speed and wind pressure of the non-building part 3b.

[0031] The machine learning device 24 performs machine learning so that the learning estimation result 16 has values ​​close to the wind speed distribution data 5 and wind pressure distribution data 6 corresponding to the input learning input data 14, i.e., the teacher data. To achieve this, the learning unit 20 compares, on a voxel-by-voxel basis, the wind speed distribution X-component data 10 of the wind speed distribution data 5 corresponding to the learning input data 14 with a first channel storing an estimation result for the wind speed in the X direction of the learning estimation result 16 obtained by inputting the learning input data 14. The learning unit 20 also compares, on a voxel-by-voxel basis, the wind speed distribution Y-component data 11 of the wind speed distribution data 5 corresponding to the learning input data 14 with a second channel storing an estimation result for the wind speed in the Y direction of the learning estimation result 16 obtained by inputting the learning input data 14. Next, the learning unit 20 compares, on a voxel-by-voxel basis, the wind speed distribution Z-component data 12 of the wind speed distribution data 5 corresponding to the learning input data 14 with a third channel storing an estimation result related to the wind speed in the Z direction of the learning estimation result 16 obtained by inputting the learning input data 14. Furthermore, the learning unit 20 compares, on a voxel-by-voxel basis, the wind pressure distribution data 6 corresponding to the learning input data 14 with a fourth channel storing an estimation result related to the wind pressure of the learning estimation result 16 obtained by inputting the learning input data 14. Then, the learning unit 20 calculates, for example, the squared error of the difference between the values ​​of each channel between each voxel obtained as a result of these comparisons as a cost function. Then, the machine learning device 24 performs machine learning by adjusting the weight values ​​of each filter using backpropagation or the like to reduce this cost function. As a result, when learning input data 14 is input, the machine learning device 24 is trained to output learning output data 15 that is close to the wind speed distribution data 5 and wind pressure distribution data 6 corresponding to the building information data 3 and wind direction data 4 contained therein. In this way, the machine learning device 24 (trained model 25 at the time of learning) performs deep learning by comparing the estimated wind speed distribution and wind pressure distribution, i.e., the learning output data 15, with the wind speed distribution data 5 and wind pressure distribution data 6 to obtain the learning estimation result 16, which is the result of multiplying the value at each position in three-dimensional space by the value of the building information data 3 corresponding to that position.

[0032] When learning is completed, the learning unit 20 stores the adjusted parameters, such as the weight values ​​of each filter, as learned model parameters in the learned model parameter storage unit 22. The learned model parameters stored in the learned model parameter storage unit 22 are acquired by the wind speed and wind pressure estimation unit 21, which will be described later, and a learned model 25 that estimates wind speed distribution and wind pressure distribution is constructed. That is, the learning unit 20 generates a trained model 25 that is used as a program module that is part of artificial intelligence software, and that has been trained with appropriate learning parameters and has completed training.

[0033] Next, the behavior of each component when estimating the wind speed distribution will be described. The wind speed and pressure estimation unit 21 acquires trained model parameters from the trained model parameter storage unit 22 and constructs a trained model 25. Figure 9 is a schematic explanatory diagram of the trained model 25. The wind speed and pressure estimation unit 21 estimates the wind speed distribution and wind pressure distribution by executing this trained model 25 as a program on a CPU, for example. Estimation input data 30, which is input data for trained model 25, is building information data 31 and wind direction data 32, which store information on the region and wind direction for which wind speed distribution and wind pressure distribution are actually estimated. Building information data 31 is configured similarly to building information data 3 used as training data 2 during training. Wind direction data 32 includes wind direction X-component data 35 and wind direction Y-component data 36, ​​each of which is configured similarly to wind direction X-component data 8 and wind direction Y-component data 9 used as training data 2 during training. Estimation input data 30 is integrated, similar to training input data 14, by, for example, setting the values ​​of voxels corresponding to the same point in building information data 31, wind direction X-component data 35, and wind direction Y-component data 36 to be values ​​of different channels within the voxels at the corresponding positions in estimation input data 30. The value of each voxel in the estimation input data 30 is normalized in the same way as during learning.

[0034] When the wind speed and pressure estimation unit 21 inputs estimation input data 30 to the trained model 25, the trained model 25 executes convolution processing and transposed convolution processing while sequentially passing through the convolution layers 27a, 27b, and 27c and the transposed convolution layers 28c, 28b, and 28a. Finally, output data 38 corresponding to the estimation input data 30 is output from the transposed convolution layer 28a. Similar to the learning output data 15, the output data 38 is an estimation result of the wind speed distribution and wind pressure distribution in the target area expressed in the building information data 31, corresponding to the building information data 31 and wind direction data 32 in the input estimation input data 30. The output data 38 is voxel data configured to have the same size as the estimation input data 30 in each of the X, Y, and Z directions. Similar to the learning output data 15, the output data 38 is configured so that each voxel has four channels, first to fourth. The first, second, third, and fourth channels of each voxel of the output data 38 store the estimation result of the wind speed in the X direction, the estimation result of the wind speed in the Y direction, the estimation result of the wind speed in the Z direction, and the estimation result of the wind pressure at the position corresponding to the voxel, respectively. In this way, the wind speed distribution estimated by the trained model 25 (i.e., output data 38) stores wind speed information corresponding to each of multiple mutually perpendicular axial directions X, Y, and Z at each position in three-dimensional space, and the wind pressure distribution estimated by the trained model 25 (i.e., output data 38) stores wind pressure information at each position in three-dimensional space.

[0035] Similar to the learning unit 20, the wind speed and wind pressure estimation unit 21 multiplies the output data 38 by the building information data 31 for each voxel, and sets the values ​​of voxels corresponding to building parts to 0, while maintaining the values ​​of voxels corresponding to non-building parts. The wind speed and wind pressure estimation unit 21 separates and extracts, as different voxel data, wind speed and wind direction information decomposed into components corresponding to the X, Y, and Z directions, stored as the first, second, and third channels, from the results of multiplication by the building information data 31 as described above. As described above, the wind speed distribution data 5 used as training data for the machine learning device 24 is normalized to a value between 0 and 1, and the trained model 25 is trained to output values ​​close to this training data. Therefore, the values ​​in each voxel are also between 0 and 1 in the output data 38 output by the trained model 25, the result of multiplying this by the building information data 31, and the voxel data extracted from the output data. The wind speed and wind pressure estimation unit 21 converts the wind speed component values ​​in the axial directions X, Y, and Z stored in each voxel into actual wind speed values. This conversion is performed in the reverse order of the standardization of the wind speed distribution data 5. Specifically, the estimated value is multiplied by 2 and then subtracted by 1 to calculate the wind speed ratio, which is then multiplied by the reference wind speed to obtain the actual wind speed. For example, if the reference wind speed is 15 m / s and the estimated value is 0.256, multiplying this by 2 and subtracting 1 gives the wind speed ratio, which is -0.488. Multiplying this wind speed ratio by the reference wind speed of 15 m / s gives the actual wind speed, which is -7.32 m / s. Alternatively, if the estimated value is 0.755, multiplying this by 2 and subtracting 1 gives the wind speed ratio, which is 0.51, and multiplying this wind speed ratio by the reference wind speed of 15 m / s gives the actual wind speed, which is 7.65 m / s. In this way, the wind speed and pressure estimation unit 21 converts the value of each voxel of the voxel data extracted for each of the X, Y, and Z directions to generate an X-component wind speed distribution estimation result 41, a Y-component wind speed distribution estimation result 42, and a Z-component wind speed distribution estimation result 43.

[0036] The wind speed and pressure estimation unit 21 also separates and extracts, as voxel data, wind pressure information corresponding to the wind pressure stored as the fourth channel from the result of multiplication by the building information data 31 as described above. As with the wind speed distribution, the value in each voxel in the separated and extracted voxel data for the wind pressure distribution is between 0 and 1. As with the wind speed, the wind speed and pressure estimation unit 21 converts the wind pressure value stored in each voxel into an actual wind pressure value. The wind speed and pressure estimating unit 21 converts the value of each voxel of the voxel data extracted in this way regarding wind pressure, and generates a wind pressure distribution estimation result 45. The wind speed and pressure estimation unit 21 outputs the X-component wind speed distribution estimation result 41, the Y-component wind speed distribution estimation result 42, the Z-component wind speed distribution estimation result 43, and the wind pressure distribution estimation result 45 as estimation results 40 to the outside. Furthermore, the wind speed and pressure estimation unit 21 transmits the X-component wind speed distribution estimation result 41, the Y-component wind speed distribution estimation result 42, and the Z-component wind speed distribution estimation result 43 to the result synthesis unit . In this way, the wind speed and wind pressure estimation unit 21 inputs building information data 31 and wind direction data 32 for estimating the wind speed distribution and wind pressure distribution into the trained model 25, estimates the wind speed distribution and wind pressure distribution, and outputs the estimation result 40 based on this.

[0037] The result synthesis unit 23 receives the X-component wind speed distribution estimation result 41, the Y-component wind speed distribution estimation result 42, and the Z-component wind speed distribution estimation result 43 from the wind speed and pressure estimation unit 21. The result combination unit 23 combines information from the X-component wind speed distribution estimation result 41, the Y-component wind speed distribution estimation result 42, and the Z-component wind speed distribution estimation result 43 to generate a combined wind speed distribution result 44. The combined wind speed distribution result 44 is voxel data of the same size as the X-component wind speed distribution estimation result 41, the Y-component wind speed distribution estimation result 42, and the Z-component wind speed distribution estimation result 43. For each corresponding voxel in the X-component wind speed distribution estimation result 41, the Y-component wind speed distribution estimation result 42, and the Z-component wind speed distribution estimation result 43, the result combination unit 23 obtains the voxel values ​​from each of these estimation results 41, 42, and 43, i.e., the X-component value, the Y-component value, and the Z-component value of the wind speed at the point corresponding to the voxel, and combines these as vectors on a three-dimensional coordinate system formed by the three axial directions of the X, Y, and Z directions. The result synthesis unit 23 calculates the wind speed based on the magnitude of the vector generated by synthesis, and the wind direction based on the direction. The result synthesis unit 23 calculates the wind speed and wind direction information for all voxels, and combines this to generate one voxel data, i.e., the wind speed distribution synthesis result 44. The result synthesis unit 23 outputs the wind speed distribution synthesis result 44 to the outside.

[0038] Fig. 10 is an example of a wind speed distribution synthesis result 44 estimated using the above-described wind speed and wind pressure estimation device 1. Fig. 11 is an example of a wind pressure distribution estimation result 45 estimated using the above-described wind speed and wind pressure estimation device 1. In reality, the wind speed and pressure distributions are estimated as voxel data by the wind speed and pressure estimation device 1, so the results are estimated in a three-dimensional space, but in Fig. 10 the results are shown as two planes: a horizontal plane corresponding to the ground surface and a plane perpendicular to it. Also, in Fig. 11 the results are shown for the wind pressure acting on the surface of the building.

[0039] Next, a method for estimating wind speed and wind pressure using the above-described wind speed and wind pressure estimation device 1 will be described with reference to Figures 1 to 11, 12 and 13. Figure 12 is a flowchart for learning wind speed distribution and wind pressure distribution, and Figure 13 is a flowchart for estimating wind speed distribution and wind pressure distribution. First, the operation of each component of the wind speed and wind pressure estimation device 1 when learning the wind speed distribution and wind pressure distribution will be described. The learning unit 20 inputs the building information data 3 and the wind direction data 4 as learning input data 14 to the convolution processing unit 27 of the machine learning device 24 . The convolution processing unit 27 performs convolution processing using each of the convolution layers 27a, 27b, and 27c to generate a feature map, and inputs the feature map to the transposed convolution processing unit . The transposed convolution processing unit 28 performs transposed convolution processing so as to enlarge and restore the feature map using each of the transposed convolution layers 28c, 28b, and 28a, and outputs the training output data 15. The learning unit 20 acquires wind speed distribution data 5 and wind pressure distribution data 6 corresponding to the input learning input data 14, and uses the wind speed distribution data 5 and wind pressure distribution data 6 as training data to compare the training data with the learning output data 15 on a voxel-by-voxel basis, and calculates, for example, the squared error of the difference in values ​​between each voxel as a cost function. Then, the machine learning device 24 performs machine learning by adjusting the weight values ​​of each filter using backpropagation or the like so as to reduce this cost function (step S1). When the learning is completed, the learning unit 20 stores the adjusted parameters such as the weight values ​​of each filter as learned model parameters in the learned model parameter storage unit 22 (step S3).

[0040] Next, the behavior of each component when estimating the wind speed distribution and wind pressure distribution will be described. The wind speed and pressure estimation unit 21 acquires the trained model parameters from the trained model parameter storage unit 22 and constructs the trained model 25 (step S11). The wind speed and pressure estimation unit 21 inputs estimation input data 30, in which the values ​​of each voxel have been normalized, to the trained model 25. The trained model 25 performs convolution processing and transposed convolution processing by sequentially passing through convolution layers 27a, 27b, and 27c and transposed convolution layers 28c, 28b, and 28a. Finally, output data 38 corresponding to the estimation input data 30 is output from the transposed convolution layer 28a. The wind speed and wind pressure estimation unit 21 multiplies the output data 38 by the building information data 31 for each voxel, and sets the value of voxels corresponding to building parts to 0, while setting the value of voxels corresponding to non-building parts to maintain the same value. The wind speed and wind pressure estimation unit 21 separates and extracts the wind speed and wind direction information, which has been decomposed into components corresponding to the X, Y, and Z directions, from the result of multiplication by the building information data 31 as described above, as different voxel data, and converts the value of each voxel into an actual wind speed value to generate an X-component wind speed distribution estimation result 41, a Y-component wind speed distribution estimation result 42, and a Z-component wind speed distribution estimation result 43. In addition, the wind speed and wind pressure estimation unit 21 separates and extracts wind pressure information corresponding to wind pressure as voxel data from the result of multiplication by the building information data 31 as described above, converts the value of each voxel to an actual wind pressure value, and generates a wind pressure distribution estimation result 45. The wind speed and pressure estimation unit 21 outputs the X-component wind speed distribution estimation result 41, the Y-component wind speed distribution estimation result 42, the Z-component wind speed distribution estimation result 43, and the wind pressure distribution estimation result 45 to the outside as estimation results 40 (step S13).

[0041] Furthermore, the wind speed and pressure estimation unit 21 transmits the X-component wind speed distribution estimation result 41, the Y-component wind speed distribution estimation result 42, and the Z-component wind speed distribution estimation result 43 to the result synthesis unit . The result synthesis unit 23 receives the X-component wind speed distribution estimation result 41, the Y-component wind speed distribution estimation result 42, and the Z-component wind speed distribution estimation result 43 from the wind speed and pressure estimation unit 21. The result synthesis unit 23 synthesizes the information of the X-component wind speed distribution estimation result 41, the Y-component wind speed distribution estimation result 42, and the Z-component wind speed distribution estimation result 43 to generate a wind speed distribution synthesis result 44 (step S15). The result synthesis unit 23 outputs the wind speed distribution synthesis result 44 to the outside.

[0042] The wind speed and pressure estimation device 1 as described above is a wind speed and pressure estimation device 1 that estimates wind speed distribution and wind pressure distribution around a building, and is provided with a trained model 25 that has been deep-trained using building information data 3 that represents the three-dimensional shape of a building and its surroundings, wind direction data 4, and wind speed distribution data 5 and wind pressure distribution data 6 in three-dimensional space as training data corresponding to the building information data 3 and wind direction data 4 as training data 2, and the trained model 25 is provided with building information data 31 and wind direction data 4 that are targets for estimating wind speed distribution and wind pressure distribution. The wind speed and pressure estimation unit 21 receives input data 32 to estimate wind speed distribution and wind pressure distribution (output data 38), and outputs an estimation result 40 based on this data. Each of the wind speed distribution data 5 and the wind speed distribution (output data 38) estimated by the trained model 25 stores wind speed information corresponding to each of multiple mutually perpendicular axial directions X, Y, and Z at each position in three-dimensional space, and each of the wind pressure distribution data 6 and the wind pressure distribution (output data 38) estimated by the trained model 25 stores wind pressure information at each position in three-dimensional space. According to the above configuration, the trained model 25 is deep-trained using the building information data 3, the wind direction data 4, and the wind speed distribution data 5 and wind pressure distribution data 6 in three-dimensional space as training data corresponding to the building information data 3 and the wind direction data 4 as training data 2. When the building information data 31 and the wind direction data 32 for which the wind speed distribution and the wind pressure distribution are to be estimated are input to the trained model 25, the trained model 25 estimates the wind speed distribution and the wind pressure distribution (output data 38). Here, the wind speed distribution data 5 and the wind speed distribution (output data 38) estimated by the trained model 25, which has undergone deep learning to output values ​​close to this wind speed distribution data 5 as training data, each store wind speed information corresponding to each of multiple mutually orthogonal axial directions X, Y, and Z at each position in three-dimensional space. That is, the trained model 25 can estimate wind speeds corresponding to multiple axial directions X, Y, and Z at each position in three-dimensional space. Furthermore, by combining the wind speeds in each of the multiple axial directions X, Y, and Z at each position, it is possible to calculate not only the wind speed but also the wind direction. Furthermore, wind pressure information at each position in three-dimensional space is stored in each of the wind pressure distribution data 6 and the wind pressure distribution (output data 38) estimated by the trained model 25, which has undergone deep learning to output values ​​close to this wind pressure distribution data 6 as training data. That is, the trained model 25 can estimate the wind pressure at each position in three-dimensional space. In particular, the trained model 25 configured as described above collectively estimates the wind speed distribution and wind pressure distribution (output data 38) at each position in three-dimensional space. That is, the trained model 25 can estimate results with continuity in each of the axial directions X, Y, and Z in three-dimensional space. Therefore, the estimation result 40 is more accurate and detailed than a method that estimates two-dimensional information. Furthermore, in the above configuration, wind speed distribution and wind pressure distribution are estimated using a deep-learned trained model 25, which can be achieved with shorter and simpler processing time than when using wind tunnel experiments or computational fluid analysis. Therefore, it is possible to realize a wind speed and wind pressure estimation device 1 that can estimate wind speed distribution and wind pressure distribution easily and in detail in a short time.

[0043] In addition, the building information data 3, 31 are modeled as three-dimensional data consisting of two values, with the building portion 3a being 0 and the non-building portion 3b being 1, and the portions of the wind speed distribution data 5 and wind pressure distribution data 6 corresponding to the building portion 3a are set to a value of 0, and during learning, the trained model 25 (machine learning device 24) performs deep learning by comparing the results (training estimation results 16) of multiplying the values ​​of the estimated wind speed distribution and wind pressure distribution (learning output data 15) at each position in three-dimensional space by the value of the building information data 3 corresponding to that position with the wind speed distribution data 5 and wind pressure distribution data 6. During learning, the trained model 25 (machine learning device 24) basically performs deep learning by comparing each of the estimated wind speed distribution and wind pressure distribution with the wind speed distribution data 5 and wind pressure distribution data 6 as training data, and reflecting the differences. In the above configuration, the values ​​of the wind speed distribution data 5 and the wind pressure distribution data 6 corresponding to the building portion 3a are set to 0. The building information data 3 is modeled as three-dimensional data consisting of two values, with the building portion 3a being 0 and the non-building portion 3b being 1. During learning, the trained model 25 (machine learning device 24) performs deep learning by comparing the results (training estimation results 16) of multiplying the values ​​of the estimated wind speed distribution and wind pressure distribution (learning output data 15) at each position in three-dimensional space by the values ​​of the building information data 3 corresponding to those positions with the wind speed distribution data 5 and the wind pressure distribution data 6. That is, in the results (training estimation results 16) of multiplying the wind speed distribution and wind pressure distribution estimated during learning (learning output data 15) by the values ​​of the building information data 3, which are the basis of deep learning and the objects for which the difference is found by comparison, and in the wind speed distribution data 5 and the wind pressure distribution data 6, both the building portion 3a is set to 0. Therefore, even if the trained model 25 (machine learning device 24) estimates a value other than 0 for the wind speed or wind pressure for the building part 3a during training, when this is compared with the training data, i.e., the wind speed distribution data 5 and the wind pressure distribution data 6, the value is multiplied by 0 to become 0, so no difference occurs with the training data, and as a result, it is not reflected in the training of the trained model 25 (machine learning device 24). With this configuration, the trained model 25 (machine learning device 24) does not need to learn excessively features related to the shape of the building, etc., and as a result, it can concentrate on learning only the features related to the wind speed and wind pressure of the non-building portion 3b. This improves the learning efficiency of the trained model 25 (machine learning device 24) and further increases the estimation accuracy of the wind speed distribution and wind pressure distribution.

[0044] Furthermore, the wind speed and wind pressure estimation method as described above is a wind speed and wind pressure estimation method for estimating wind speed distribution and wind pressure distribution around a building, and includes a step of inputting building information data 3 representing the three-dimensional shape of a building and its surroundings, wind direction data 4, and wind speed distribution data 5 and wind pressure distribution data 6 in three-dimensional space as teacher data corresponding to the building information data 3 and wind direction data 4 as learning data 2 into a trained model 25 that has been deep-learned, to estimate wind speed distribution and wind pressure distribution (output data 38), and outputting estimation results 40 based on this. Each of the wind speed distribution data 5 and the wind speed distribution (output data 38) estimated by the trained model 25 stores wind speed information corresponding to each of multiple mutually perpendicular axial directions X, Y, and Z at each position in the three-dimensional space, and each of the wind pressure distribution data 6 and the wind pressure distribution (output data 38) estimated by the trained model 25 stores wind pressure information at each position in the three-dimensional space. According to the above configuration, it is possible to realize a wind speed and wind pressure estimation method that can estimate wind speed distribution and wind pressure distribution easily and in detail in a short time.

[0045] In the wind speed and pressure estimation device 1 and wind speed distribution estimation method shown as the above embodiment, the more training data 2 is prepared and used for training the machine learning device 24, the more likely it is that the estimation accuracy of the trained model 25 will improve. A method for realizing such a large amount of training data 2 will be described next. For example, there may be cases where the area represented in the building information data 3, wind speed distribution data 5, and wind pressure distribution data 6 is too large to be suitable for learning, or is larger than expected for the actual operation of the wind speed and pressure estimation device 1. In such cases, multiple pieces of learning data 2 can be generated from one piece of learning data 2 by extracting an appropriate range from any part of the large area for each piece of voxel data relating to the building information data 3, wind speed distribution data 5, and wind pressure distribution data 6. In this case, it is necessary to extract the building information data 3, wind speed distribution data 5, and wind pressure distribution data 6 at the same position and range, and to ensure that voxels located at the same coordinates in each voxel data correspond to the same point. Similarly, new training data 2 can also be generated by rotating each voxel data in the training data 2 around an arbitrary rotation axis parallel to the Z direction. New learning data 2 can also be generated by inverting each voxel data in the learning data 2 with respect to an arbitrary plane or axis parallel to the Z direction. In this way, by generating multiple pieces of training data 2 from one piece of training data 2, a large amount of training data 2 can be used for training the machine learning device 24, thereby improving the estimation accuracy of the trained model 25.

[0046] Such generation of learning data 2 from one learning data 2 may be performed by randomly setting the cut-out position, range, rotation axis, rotation direction, rotation angle, inversion plane, etc. for the learning data 2. However, for example, at least the ground surface must be included in the voxel data, and there is a possibility that inappropriate data will be generated if the range is randomly cut out in the Z direction. Similarly, if the earth's surface is positioned at the top of the voxel data and the data is inverted, inappropriate data will be generated. For this reason, when randomly generating the learning data 2, it is necessary to appropriately set the cut-out position and range, rotation axis, rotation direction, rotation angle, inversion plane, etc.

[0047] The wind speed distribution estimation device and wind speed distribution estimation method of the present invention are not limited to the above-described embodiment explained with reference to the drawings, and various other modifications are conceivable within the technical scope thereof. For example, in the above embodiment and modified examples, an FCN is used as the machine learning device 24 and the trained model 25, but neural networks with other structures may be used as long as the estimation accuracy is not impaired. 2 and 9, the structure of the FCN may be other than the above. For example, in the above embodiment, the convolution processing unit 27 and the transposed convolution processing unit 28 are each described as having a three-layer structure as a schematic example, but the number of layers of each may be other than three, or the number of layers of the convolution processing unit 27 and the number of layers of the transposed convolution processing unit 28 may be different. Furthermore, in the above embodiment, the wind direction data 4 was data representing the direction of wind blowing from outside the area toward a building or group of buildings within the area represented in the building information data 3, but the wind direction data 4 may also include information on wind speed in addition to wind direction.

[0048] Furthermore, in the above embodiment, in order to improve learning efficiency by concentrating only on the features related to wind speed and wind pressure in non-building portion 3b and training trained model 25 (machine learning device 24), during learning, learning output data 15 was multiplied by the value of building information data 3 to generate learning estimation result 16, which was then compared with the teacher data. Correspondingly, during estimation of wind speed distribution and wind pressure distribution, output data 38 was multiplied by building information data 31, and estimation result 40 was generated based on this result. However, if sufficient learning efficiency can be obtained, it is not necessary to perform the multiplication of these building information data 3 and 31. In other words, in this case, the learning output data 15 may be directly compared with the teacher data during learning. Furthermore, when estimating the wind speed distribution and wind pressure distribution, the output data 38 may be directly output as the estimation result 40. In addition to this, it is possible to select and discard the configurations given in the above embodiments and modifications, or to change them to other configurations as appropriate, without departing from the spirit of the present invention. [Explanation of symbols]

[0049] 1 Wind speed and pressure estimation device 20 Learning unit 2 Training data 21 Wind speed and pressure estimation part 3, 31 Building information data 23 Result synthesis section 3a Building part 24 Machine learning machine (trained model during training) 3b Non-building area 25 trained model 4, 32 Wind direction data 30 Input data for estimation 5 Wind speed distribution data 38 Output data 6 Wind pressure distribution data 40 Estimation results 14 Input data during learning 44 Wind speed distribution synthesis results 15 Learning output data X, Y, Z axis directions 16 Estimation results during training

Claims

1. A wind speed and pressure estimation device that estimates wind speed distribution and wind pressure distribution around a building, a trained model that has been deep-trained using, as training data, building information data that represents the three-dimensional shape of a building and its surroundings using voxel data in which voxels, the smallest three-dimensional units, are connected; wind direction data that represents the wind direction using voxel data in which component values ​​in mutually orthogonal axial directions are set in the voxels; and wind speed distribution data and wind pressure distribution data in three-dimensional space that serve as teacher data corresponding to the building information data and the wind direction data; and a wind speed and wind pressure estimation unit that inputs the building information data and wind direction data that are the targets for estimating the wind speed distribution and the wind pressure distribution into the trained model, estimates the wind speed distribution and the wind pressure distribution, and outputs an estimation result based on this; The wind speed distribution data and the wind speed distribution estimated by the trained model are each represented by voxel data in which a value indicating a vector representing the wind direction and wind speed at a point corresponding to each voxel is stored in the voxel, and the wind pressure distribution data and the wind pressure distribution estimated by the trained model are each represented by voxel data in which a value indicating the wind pressure at a point corresponding to each voxel is stored in the voxel, A wind speed and pressure estimation device in which the wind speed distribution data has values ​​normalized by wind speed at or above a boundary layer height, which indicates the height of the layer at which wind speed attenuates, depending on the size or density of the buildings, and the wind pressure distribution data has values ​​normalized by a reference wind pressure at or above the boundary layer height.

2. The building information data is modeled as voxel data consisting of two values, with the value of voxels of building parts being 0 and the value of voxels of non-building parts being 1, the wind speed distribution data and the wind pressure distribution data are voxel data in which the value of a voxel in a portion corresponding to the building portion is set to 0, The wind speed and wind pressure estimation device of claim 1, characterized in that during learning, the trained model is deep-learned by multiplying the value of each voxel of the estimated wind speed distribution and wind pressure distribution by the value of the voxel of the building information data corresponding to that voxel, and comparing the result with the wind speed distribution data and the wind pressure distribution data.

3. A wind speed and wind pressure estimation method for estimating wind speed distribution and wind pressure distribution around a building, comprising: The method comprises a step of inputting building information data representing the three-dimensional shape of a building and its surroundings using voxel data in which voxels, the smallest three-dimensional units, are connected, wind direction data representing wind direction using voxel data in which component values ​​in mutually orthogonal axial directions are set in the voxels, and wind speed distribution data and wind pressure distribution data in three-dimensional space as teacher data corresponding to the building information data and the wind direction data into a trained model that has been trained using deep learning as training data, the building information data and wind direction data from which the wind speed distribution and the wind pressure distribution are to be estimated, estimating the wind speed distribution and the wind pressure distribution, and outputting an estimation result based on the input data; The wind speed distribution data and the wind speed distribution estimated by the trained model are each represented by voxel data in which a value indicating a vector representing the wind direction and wind speed at a point corresponding to each voxel is stored in the voxel, and the wind pressure distribution data and the wind pressure distribution estimated by the trained model are each represented by voxel data in which a value indicating the wind pressure at a point corresponding to each voxel is stored in the voxel, A method for estimating wind speed and wind pressure, wherein the wind speed distribution data has values ​​normalized by wind speed at or above a boundary layer height, which indicates the height of the layer at which wind speed attenuates, depending on the size or density of the buildings, and the wind pressure distribution data has values ​​normalized by a reference wind pressure at or above the boundary layer height.

Images